Pair-diff eval

Pairs, not scores.

Run the same prompts against your incumbent and Anomalia. We sort by divergence so you see the twenty differences that matter. Scores invite arguments; ordering does not.


Demo mode loads sample CLI-style prompts. Edit freely, then run.

Keys stay in your browser. Calls go from your machine to the endpoints you named. CORS must allow this origin — if blocked, use a local proxy.

Sorted by divergence

No run yet

Run the eval to see side-by-side pairs, highest divergence first.